What HarnessHealth actually does
AI can draft clinical work in seconds — notes, letters, care plans, billing codes. HarnessHealth is the layer that makes sure a licensed physician signs that work before it counts.
That is the whole idea. Everything else on this site is detail. This page walks the idea end to end, then hands you three things you can try right now without signing up.
The problem, in one story
This is happening today, in thousands of tools, with no gate in the middle.
An AI drafts it
A care plan, a prior auth, a letter of medical necessity. It takes seconds and it reads perfectly — that is exactly the problem. Fluency is not accuracy.
Why fluent output still needs governance →Nobody signs it
No NPI attached. No named reviewer. If the draft is wrong and a patient acts on it, whose license is on the line? Today the answer is: nobody’s.
The governance model →It reaches a patient anyway
Most AI health tools ship the draft with a disclaimer. A disclaimer is not a gate. This is the gap between capability and accountability.
How the gap gets closed →The harness, in three moves
Draft → signed → shipped. Each move is enforced by architecture, not by policy.
AI drafts — and the draft is held
Every AI output starts life as a DRAFT that physically cannot leave the system. Not a policy, not a setting: there is no code path that transmits an unsigned draft to a patient, a payer, or a medical record.
The hard intercept, in detail →A physician signs — and the signature is sealed
A licensed physician reviews the draft, typically in 3–5 minutes. Their NPI, a cryptographic fingerprint of the exact document, and a timestamp are locked together in a record that cannot be edited afterward.
What review looks like for a physician →Only then it ships — and it stays provable
The output becomes ATTESTED: billable to the codes it supports, defensible in an audit, and verifiable years later. The review pace is itself monitored, so a signature can never become a rubber stamp.
The published evidence for this architecture →Approve is a click. Attest is a record.
The model is becoming a commodity — GPT, Claude, an open model, whichever you prefer. When the intelligence is interchangeable, trust is the product. That is the layer HarnessHealth builds, and it is not one guarantee but two.
Anyone can bolt an “approve” button onto an AI output. That proves a button was pressed. Attestation answers the two questions a button can’t — and both answers are permanent records, not settings someone can flip later.
Who stands behind this?
A named, licensed physician reviewed the exact document and bound their NPI, a cryptographic fingerprint, and a timestamp to it. Not “a doctor was in the loop” — this doctor, on this text, at this moment, provably.
How the signature is sealed →Where did the private data go?
The other half of trust: a record of where the protected data didn’t go — not into a URL, a log, an analytics pixel, or a model’s training set. A signed output on top of leaked PHI is not a safe system. Both records travel together.
Who owns the conversation? →See it work, right now
Three things on this site run live today. None of them ask for an email.
The Fidelity Gate
A real care plan, an AI summary that quietly drops the penicillin allergy, and a gate that blocks it. Edit the summary yourself — it re-scores as you type. Deterministic, so you can trust what you see.
Open the gate demo →Governance self-check
Answer a few questions about any AI deployment you run or are evaluating — yours or a vendor’s — and get a governance score you can put in front of a compliance officer.
Score a deployment →Find a physician
Every U.S. physician in the CMS registry — 2.4 million of them — already has a page here, no signup required. Look yourself up, or find the specialist you refer to.
Search the directory →What you get from it
The same gate, three different reasons to care.
Your license becomes an asset that earns while you sleep.
AI does the drafting; you do the judging. Reviews take 3–5 minutes each, your NPI is bound to every one, and the earnings calculator shows what a realistic review load pays.
See the earnings calculator →Deploy AI without inheriting the liability question.
The harness sits in front of any model your teams already use — Claude, GPT, or your own — and guarantees nothing clinically meaningful ships unsigned. BAA available.
The health-system deployment path →Add a physician network instead of recruiting one.
Your app generates the clinical output; the harness routes it to a licensed reviewer and returns it signed. The SDK pipes — FHIR, HIPAA transport, PROM tools — are free.
Developer integration guide →What is live today — and what is still being built
Trust infrastructure should not overstate itself. Here is the honest line.
- The Fidelity Gate — deterministic, edit-it-yourself demo of governed output
- Governance self-check — score any AI deployment against the harness standard
- The physician directory — 2.4M NPI pages, pre-indexed from CMS
- Billing integrity inside SurgeonValue — NCCI correct-coding validation, a deterministic rules engine with no LLM in the loop
- Public attestation API — the endpoints named in the agent card are in development; integration today runs through a partnership conversation
- Published validation — the peer-reviewed existence proofs are live; our own outcome data accrues as the network reviews
- Health-system pilots — deployment path is defined; early conversations are open now
Ninety seconds is up.
The fastest way to understand the harness is to watch it block something. The second fastest is a thirty-minute conversation.